Ed Baker was exactly the kind of person Tesla would want behind the wheel of a Full Self-Driving vehicle. A former vice president of product and growth at Uber, Baker understood autonomous technology intimately. He’d spent years immersed in the economics and engineering of self-driving systems. He believed in the promise. And then, on a stretch of California road, his Tesla — operating on FSD — drove straight into a concrete median.
He wasn’t hurt. But the incident left a mark that went beyond body damage on a Model S.
Baker has since become one of the more credible voices raising alarms about Tesla’s approach to autonomous driving, not from the position of a skeptic who never trusted the technology, but as a true believer who watched it fail in real time. As Business Insider reported, Baker’s experience — and his subsequent public commentary — carries particular weight precisely because of his pedigree. This isn’t a technophobe railing against progress. It’s an insider who has seen both sides of the autonomous divide and is now questioning whether Tesla’s timeline and methodology are putting lives at risk.
The crash itself was, in Baker’s telling, disturbingly mundane. The car was operating on Tesla’s FSD supervised mode, which despite its name requires constant driver attention. The system misjudged the road geometry and steered into a concrete barrier. Baker, who had been monitoring the system as instructed, didn’t have enough time to override the vehicle’s decision. The car was damaged. He walked away. But what stayed with him was the gap between Tesla’s marketing language and the technology’s actual capability — a gap he now argues is dangerous.
Tesla’s Full Self-Driving software has been the subject of intense debate for years. The company charges up to $12,000 for the feature (or $199 per month as a subscription), and CEO Elon Musk has repeatedly promised that fully autonomous driving is imminent. In 2016, Musk said a Tesla would drive itself from Los Angeles to New York by 2017. That hasn’t happened. In 2020, he said FSD would be “feature complete” by year’s end. The goalposts have moved so many times that even bullish analysts have stopped penciling in specific dates.
And yet the stakes keep rising.
Tesla launched its robotaxi service ambitions with the unveiling of the Cybercab in late 2024, a vehicle with no steering wheel and no pedals, designed exclusively for autonomous operation. Musk has said Tesla plans to begin robotaxi operations in Austin, Texas, in June 2025, initially using Model 3 and Model Y vehicles with human safety drivers, before eventually transitioning to fully driverless Cybercabs. The company’s entire growth narrative now hinges on autonomy. Morgan Stanley analyst Adam Jonas has valued Tesla’s robotaxi business at over $500 billion in his most optimistic scenarios. If FSD doesn’t work — really work, not demo-reel work — that valuation collapses.
Baker’s concern isn’t that autonomy will never arrive. It’s that Tesla’s approach — relying solely on cameras and neural networks without lidar or radar redundancy — creates failure modes that are difficult to predict and nearly impossible for a human supervisor to catch in time. As he explained to Business Insider, the problem with AI-driven systems is that they don’t fail the way humans expect. A human driver who’s confused slows down. An AI system that’s confused might accelerate into a median with full confidence. The system doesn’t know what it doesn’t know.
This is not a theoretical concern. The National Highway Traffic Safety Administration has opened multiple investigations into Tesla’s Autopilot and FSD systems. NHTSA’s Standing General Order data, which requires manufacturers to report crashes involving automated driving systems, shows hundreds of incidents involving Tesla vehicles. In several cases, the crashes resulted in fatalities. Tesla has issued recalls — delivered as over-the-air software updates — but critics argue that software patches don’t address fundamental architectural limitations.
The broader self-driving industry has largely converged on a different approach. Waymo, the Alphabet subsidiary that operates commercial robotaxi services in San Francisco, Phoenix, Los Angeles, and Austin, uses a sensor suite that includes cameras, lidar, and radar. So does Cruise, General Motors’ autonomous vehicle unit, which suspended operations in late 2023 after an incident in San Francisco but has been working toward a relaunch. Zoox, Amazon’s self-driving subsidiary, also uses a multi-sensor approach. The industry consensus is clear: redundancy saves lives. Tesla disagrees.
Musk has long argued that lidar is a “crutch” and that a vision-only system, modeled on how humans drive using just their eyes, is the correct long-term solution. There’s an elegant logic to this. Humans manage to drive with two eyes and a brain. Why can’t a sufficiently advanced neural network do the same with eight cameras? The answer, as Baker and others have pointed out, is that the comparison breaks down under scrutiny. Human vision operates with depth perception, peripheral awareness, and a lifetime of contextual learning that no current AI system can match. Humans also drive poorly — roughly 40,000 people die on American roads every year — which is an odd benchmark for a technology that’s supposed to be safer.
The financial implications of this debate are enormous. Tesla’s stock price, which has swung wildly over the past two years, is heavily influenced by investor belief in the autonomy story. When Musk announced the Cybercab and robotaxi plans, the stock surged. When FSD demonstrations have gone poorly — or when regulatory scrutiny has intensified — it’s dropped. The company’s market capitalization, which has at times exceeded $1 trillion, prices in a future where Tesla isn’t just a car manufacturer but a transportation platform. Strip away the autonomy premium and Tesla trades at a dramatically different multiple.
Baker’s warning resonates because it comes at a moment of acceleration. Tesla is pushing to deploy FSD more broadly, rolling out the software to more vehicles and in more markets. The company recently began offering FSD trials to all new Tesla buyers, a move designed to increase adoption and generate the driving data that Tesla’s neural networks need to improve. More miles driven on FSD means more training data. But it also means more exposure to edge cases — the unusual, unpredictable situations where AI systems are most likely to fail.
Edge cases are the unsolved problem of autonomous driving. A plastic bag blowing across a highway. A mattress fallen from a truck. A child chasing a ball into the street from behind a parked van. These are scenarios that human drivers handle through intuition, experience, and split-second judgment. AI systems handle them through pattern matching against training data. If the system has never seen a particular scenario — or has seen it too few times to generalize — it may respond incorrectly. Or not at all.
Baker’s crash was, in a sense, an edge case. The road geometry confused the system. A human driver would have recognized the concrete median and steered around it. The AI didn’t. And Baker, despite being an experienced technology executive who understood the system’s limitations, couldn’t intervene quickly enough. This is the fundamental paradox of supervised autonomy: the better the system works most of the time, the less prepared the human supervisor is to take over when it fails. Psychologists call this the automation complacency problem. It’s well-documented in aviation, where pilots who rely heavily on autopilot systems sometimes struggle to respond when those systems disengage unexpectedly.
Tesla’s response to these criticisms has been consistent. The company points to data suggesting that vehicles operating on Autopilot or FSD are involved in fewer crashes per mile than the average American driver. Musk has called FSD “the safest way to drive” and has accused critics of ignoring the statistical evidence. But independent researchers have questioned Tesla’s methodology, noting that FSD is primarily used on highways and in good weather — conditions where crash rates are already lower regardless of what technology is involved. Comparing FSD’s highway performance to the national average, which includes drunk driving, rural roads, and adverse weather, is not an apples-to-apples analysis.
The regulatory picture is fragmented and evolving. California, where Baker’s crash occurred, requires companies testing autonomous vehicles to report all crashes and disengagements to the state’s Department of Motor Vehicles. Tesla has historically resisted this reporting framework, arguing that FSD is a driver-assistance system rather than an autonomous vehicle technology — a distinction that critics call a semantic dodge designed to avoid regulatory oversight. Other states have taken a more permissive approach. Texas, where Tesla plans to launch its robotaxi service, has relatively light autonomous vehicle regulations, which is one reason the company chose Austin as its initial market.
The tension between innovation and safety is not new. Every transformative technology — aviation, nuclear power, the internet — has gone through a period where the pace of development outstripped the capacity of regulators and the public to assess risk. But autonomous vehicles present a unique challenge because the consequences of failure are immediate, physical, and deeply personal. A software bug in a social media algorithm might spread misinformation. A software bug in a self-driving car might kill someone.
Baker has been careful to frame his criticism constructively. He’s not calling for a ban on autonomous driving technology. He’s arguing for greater transparency, more rigorous testing, and a regulatory framework that holds companies accountable for the performance of their systems in real-world conditions. He’s also questioning whether Tesla’s business model — which relies on selling FSD to individual consumers rather than operating a managed fleet — creates perverse incentives. A fleet operator like Waymo can control the conditions under which its vehicles operate, restricting service during heavy rain or in unfamiliar areas. A Tesla owner with FSD can activate the system anywhere, anytime, in any conditions.
That distinction matters more than it might seem.
Waymo’s vehicles have now completed millions of autonomous miles in their operating territories. The company publishes detailed safety reports and works closely with regulators. Its vehicles still make mistakes — no autonomous system is perfect — but the controlled deployment model means that errors can be identified, analyzed, and corrected before the system is expanded to new areas. Tesla’s approach is the opposite: deploy broadly, collect data, and iterate. It’s a model that has worked spectacularly well for consumer software. Whether it works for two-ton vehicles traveling at highway speeds is a different question entirely.
The AI risk dimension of Baker’s argument deserves particular attention. As artificial intelligence systems become more capable and more widely deployed, the question of how they fail — and who bears responsibility when they do — is becoming one of the defining issues of the decade. Autonomous vehicles are in many ways the most visible test case. They operate in public spaces, interact with vulnerable road users, and make decisions with life-or-death consequences in fractions of a second. How society handles the risks of self-driving cars will set precedents for how it handles AI risk in healthcare, finance, criminal justice, and national security.
Baker understands this. His background at Uber, where the company’s autonomous vehicle program was suspended after a fatal crash in Tempe, Arizona, in 2018, gives him a perspective that few other commentators share. That crash — in which an Uber test vehicle struck and killed pedestrian Elaine Herzberg — was a watershed moment for the industry. It demonstrated that autonomous vehicle failures aren’t abstract. They’re fatal. And they carry legal, ethical, and reputational consequences that can set an entire technology back by years.
Tesla has so far avoided a single defining incident of that magnitude. But the accumulation of smaller incidents — crashes, near-misses, NHTSA investigations, recall after recall — is building a pattern that regulators and plaintiffs’ attorneys are watching closely. If Tesla’s robotaxi service launches in Austin this summer and a serious accident occurs, the fallout could be severe. Not just for Tesla, but for the entire autonomous vehicle industry.
Musk, characteristically, is undeterred. He has described FSD as “insanely great” and has predicted that Tesla’s autonomous technology will eventually be so safe that regulators will have no choice but to approve fully driverless operation everywhere. He may be right. The technology is improving rapidly, and the version of FSD available today is substantially better than what was available two years ago. But “substantially better” and “safe enough to operate without human supervision” are very different standards. And the distance between them may be larger than Tesla’s timeline suggests.
Baker’s crash didn’t make national headlines when it happened. It was one incident among many, a data point in a growing dataset of FSD failures. But his willingness to speak publicly about it — and to connect it to broader questions about AI safety, corporate responsibility, and regulatory adequacy — has given the incident an outsized significance. When a former Uber executive who understands autonomous technology at a granular level says the system isn’t ready, it’s harder to dismiss than when the same criticism comes from someone who’s never used the technology.
The months ahead will be telling. Tesla’s Austin robotaxi launch, if it happens on schedule, will be the most significant test of the company’s autonomous technology to date. Waymo continues to expand, recently announcing plans for service in Atlanta and Miami. Regulatory agencies at both the state and federal level are drafting new rules for autonomous vehicles. And the broader AI safety conversation — fueled by rapid advances in large language models and generative AI — is creating a political environment where calls for oversight are gaining traction across party lines.
Ed Baker drove his Tesla into a median and walked away with a story. The question now is whether the industry — and the regulators charged with overseeing it — will listen before the next story doesn’t end as well.
WebProNews is an iEntry Publication